The TensorMol-0.1 model chemistry: a neural network augmented with long-range physics.

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Traditional force fields lack accuracy for chemical reactivity and require frequent re-fitting.
• Neural network potentials (NNPs) offer near *ab initio* accuracy at a lower computational cost.
• Data-driven NNPs are inefficient for modeling long-range interatomic forces.

Purpose of the Study:

• To develop a hybrid model chemistry combining NNPs with physical models for improved accuracy and efficiency.
• To introduce TensorMol-0.1, an open-source Python package for molecular simulations.
• To demonstrate the robustness, speed, and accuracy of the developed model.

Main Methods:

• Constructed a hybrid model using a nearsighted NNP with screened long-range electrostatic and van der Waals physics.
• Implemented the model in an open-source Python package supporting various simulation types.
• Validated the model's accuracy and scalability through benchmark calculations and simulations.

Main Results:

• Achieved millihartree accuracy, comparable to electronic structure theory.
• Demonstrated scalability to tens-of-thousands of atoms on standard hardware.
• Successfully reproduced vibrational spectra and simulated protein molecular dynamics.
• Showcased the package's capability for geometry optimizations, molecular dynamics, and Monte Carlo simulations.

Conclusions:

• The hybrid NNP model significantly enhances molecular simulation capabilities.
• TensorMol-0.1 lowers the resource barrier for complex chemical simulations.
• Neural network molecular dynamics is emerging as a powerful tool for molecular simulation, bridging accuracy and efficiency gaps.